Interpreting Open-Domain Modifiers: Decomposition of Wikipedia Categories into Disambiguated Property-Value Pairs

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6218–6228, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 6218 Interpreting Open-Domain Modifiers: Decomposition of Wikipedia Categories into Disambiguated Property-Value Pairs Marius Pas¸ca Google 1600 Amphitheatre Parkway Mountain View, California 94043 mars@google.com Abstract This paper proposes an open-domain method for automatically annotating modifier con- stituents (“20th-century”) within Wikipedia categories (“20th-century male writers”) with properties (“date of birth”). The annotations offer a semantically-anchored understanding of the role of the constituents in defining the underlying meaning of the categories. In experiments over an evaluation set of Wiki- pedia categories, the proposed method an- notates constituent modifiers as semantically- anchored properties, rather than as mere strings in a previous method. It does so at a better trade-off between precision and recall. 1 Introduction Motivation: As Web search moves towards return- ing structured answers rather than flat sets of docu- ment links in response to users’ queries, the need for high-quality, wide-coverage knowledge to sup- port such answers is growing stronger. The largest of the existing knowledge repositories (Bizer et al., 2009; Hoffart et al., 2013; Nastase and Strube, 2013), whether publicly available (Bollacker et al., 2008; Vrande˘ci´c and Kr¨otzsch, 2014) or restricted to commercial access (Wu et al., 2012), uniformly rely on data in Wikipedia for their core sets of top- ics and knowledge assertions. In addition to its role in Web search and information retrieval (Chen et al., 2017; Ensan and Bagheri, 2017; Ma et al., 2018; Zhang and Balog, 2018), Wikipedia and the knowledge repositories derived from it are use- ful in a growing variety of tasks. Such tasks per- tain to text analysis (Ratinov et al., 2011; Murty et al., 2018) and, specifically, knowledge acquisi- tion from text (Nastase and Strube, 2013; Wu et al., 2012; Hoffart et al., 2013; Gupta et al., 2019). Millions of Wikipedia articles are connected to parent categories, which are in turn connected to their own, iteratively broader categories. For exam- ple, the article “art:Gary Oldman” is connected to parent categories such as “ctg:20th-century En- glish male actors”, “ctg:Alumni of Rose Bruford College”, which are in turn connected to their own, broader categories such as “ctg:Actors”, “ctg:Acting”, “ctg:Language”. Some categories re- ally correspond to individual topics, e.g., “ctg:Rose Bruford College”. Many other categories in Wiki- pedia - hundreds of thousands, by our estimates - each corresponds to a fine-grained class that groups together individual articles sharing common properties. For example, the category “ctg:20th- century English male actors” groups articles such as “art:Gary Oldman” and “art:Jude Law”, which conceptually share properties that could be de- scribed as “born or living in the 20th-century”, “born in England”, “being a male” and “being ac- tors”. That Wikipedia organizes topics into hun- dreds of thousands of potential fine-grained classes is remarkable. Comparatively, topics in other knowledge repositories are organized into only hun- dreds of types, in DBpedia (Bizer et al., 2009); or thousands of collections, in Freebase (Bollacker et al., 2008). Existing applications (Ma et al., …

Content truncated to 500 words